Use(load) the npy of the first picture to continue training the npy of the second picture about stylegan2-projecting-images

[issue-label-bot]

Issue-Label Bot is automatically applying the label question to this issue, with a confidence of 0.72. Please mark this comment with 👍 or 👎 to give our bot feedback!

Links: app homepage, dashboard and code for this bot.

[woctezuma]

It is an interesting idea. I have not done it, but I think you would only have to edit:

https://github.com/NVlabs/stylegan2/blob/dae148b9bb6346b00c10f0cfc9e011b3ef0c1986/projector.py#L160-L162

def start(self, target_images):
    [...]

    # Prepare target images.
    self._info('Preparing target images...')
    [...]

    # Initialize optimization state.
    self._info('Initializing optimization state...')
    tflib.set_vars({
                    self._target_images_var: target_images, 
                    self._dlatents_var: np.tile(self._dlatent_avg, [self._minibatch_size, 1, 1])
                    })
    [...]

so that the optimization starts from an array w fed as input by the user instead of always the same w_avg:

np.tile(self._dlatent_avg, [self._minibatch_size, 1, 1])

It is a bit similar to the idea of improving the initialization here: rolux/stylegan2encoder#2